Research map / Vision-language and multimodal / Robot action and navigation
Dita: Scaling Diffusion Transformer for Generalist Vision-Language-Action Policy
Zhi Hou, Tianyi Zhang, Yuwen Xiong, Haonan Duan, Hengjun Pu, Ronglei Tong, Chengyang Zhao, Xizhou Zhu, Yu Qiao, Jifeng Dai, Yuntao Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace diffusion action head with in-context conditioning in the DiT (ablation) | Dita · CALVIN ABC->D | average length: 3.16 -> 3.61 | Improved |
| add OXE pretraining (ablation) | Dita · CALVIN ABC->D | average length: 2.38 -> 3.61 | Improved |
| increase observation frames from 2 to 3 (ablation) | Dita · ManiSkill2 | avg success rate (%): 65.8% -> 35.4% | Hurt |
| use 2 observation frames instead of 1 (ablation) | Dita · ManiSkill2 | avg success rate (%): 61.6% -> 65.8% | Improved |
| increase trajectory length from 2 to 32 (ablation) | Dita · ManiSkill2 | avg success rate (%): 40.8% -> 65.8% | Improved |
| in-context conditioning instead of AdaLN DiT (ablation) | Dita · CALVIN ABC->D (no pretraining) | average length: 1.45 -> 2.56 | Improved |
| add learning rate decay in finetuning (ablation) | Dita · CALVIN ABC->D | average length: 3.43 -> 3.61 | Improved |
| scale transformer from 12 to 24 layers (ablation) | Dita without pretraining · CALVIN ABC->D | average length: 2.56 -> 2.61 | Improved |
| add image augmentation during finetuning (ablation) | Dita · CALVIN ABC->D | average length: 3.61 -> 3.78 | Improved |
| reduce DDIM inference steps from 20 to 2 (ablation) | Dita · SimplerEnv Pick Coke (variant) | success rate (%): 85.5 -> 70.4 | Hurt |
| reduce DDIM inference steps from 20 to 10 (ablation) | Dita · SimplerEnv Pick Coke (variant) | success rate (%): 85.5 -> 85.3 | No clear effect |
| larger shuffle buffer (256000 vs 128000) (ablation) | Dita · SimplerEnv coke_can (match) | success rate (%): 71.2% -> 83.7% | Improved |
Later papers that compared with this method
2 later accepted papers report Dita as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- PixelVLA: Advancing Pixel-level Understanding in Vision-Language-Action Model (ICLR 2026 · better on 1 of 3 shared results)
- Spatial Forcing: Implicit Spatial Representation Alignment for Vision-language-action Model (ICLR 2026 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Prediction with Action: Visual Policy Learning via Joint Denoising Process (NeurIPS 2024 · similarity 0.91)
- TDSR-VLA: Transition-aware Denoising Sequence Representations for Vision-Language-Action (ECCV 2026 · similarity 0.91)
- DiT4DiT: Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control (NeurIPS 2026 · similarity 0.90)
- Conditioning Matters: Training Diffusion Policies is Faster Than You Think (NeurIPS 2025 · similarity 0.89)
- Learning an Actionable Discrete Diffusion Policy via Large-Scale Actionless Video Pre-Training (NeurIPS 2024 · similarity 0.89)
- Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
